I’m guessing because China isn’t producing the type of RAM they want. Put it this way, when the AI bubble bursts we aren’t going to have a load of RAM and GPUs suddenly available, because the AI data centre versions aren’t in a format that’s useful in consumer tech, or even all the data centre systems. They are optimised towards AI workloads and don’t really work in normal computer systems. Equally that means that if the RAM that’s been produced isn’t in this AI workload optimised format, then the data centres can’t use it.
The crime isn’t that the AI data centres were using a lot of the RAM, it was that the manufacturers were letting them by producing all their RAM in the required format.
Without actually ripping chips off boards, you would be stuck with a pizza oven sized computer that sounds like a jet engine, uses more electricity than that pizza oven would, and doesn’t have a video-out port on the GPU. They really are tailored to a particular kind of use case and unless you are running your own ai models or /maybe/ something like crypto mining, its going to be pretty useless.
Even if they were made to work, they may still be bad in price / perf and not worth using simply due to being manufactured for a completely different purpose.
when the bubble pops there will be loads of these (assuming they haven’t all been cocktailed).
prob buy em for pennies on the dollar.
bet in 10 yrs time someone could figure out how to run them on a quarter the energy usage.
one can hope, to up cycle this fucking stain on humanity
Why aren’t data centers buying up all this RAM too?
I’m guessing because China isn’t producing the type of RAM they want. Put it this way, when the AI bubble bursts we aren’t going to have a load of RAM and GPUs suddenly available, because the AI data centre versions aren’t in a format that’s useful in consumer tech, or even all the data centre systems. They are optimised towards AI workloads and don’t really work in normal computer systems. Equally that means that if the RAM that’s been produced isn’t in this AI workload optimised format, then the data centres can’t use it.
The crime isn’t that the AI data centres were using a lot of the RAM, it was that the manufacturers were letting them by producing all their RAM in the required format.
Ehh HBM3 memory is definitey very useful in consumer GPUs. I have a prosumer one with HBM2 and it helps. The bandwidth is several times that of GDDRX!
but surely there are ppl out there smart enough to make it work somehow?
after all we can play Doom on just about any device
Without actually ripping chips off boards, you would be stuck with a pizza oven sized computer that sounds like a jet engine, uses more electricity than that pizza oven would, and doesn’t have a video-out port on the GPU. They really are tailored to a particular kind of use case and unless you are running your own ai models or /maybe/ something like crypto mining, its going to be pretty useless.
I’m sure they could be made to work. But why would they when they’ve got the likes of Nvidia producing all the chips they could ever want.
Even if they were made to work, they may still be bad in price / perf and not worth using simply due to being manufactured for a completely different purpose.
when the bubble pops there will be loads of these (assuming they haven’t all been cocktailed).
prob buy em for pennies on the dollar.
bet in 10 yrs time someone could figure out how to run them on a quarter the energy usage.
one can hope, to up cycle this fucking stain on humanity